Term-level box constraints potentially prevent convergence detection

[krivit]

Most convergence detection algorithms we use (Hotelling, Hummel, confidence) rely on simulated values of estimating functions equalling on average the observed. However, with a box constraint, such as GWESP or GWDegree decay being constraint to be >= 0, if the MLE is outside the box, this might never happen.

This will also affect #296, if implemented.

A quick and dirty mitigation might be to check which parameter vector elements are on the edge of their box and for those, "suspend" the tests for their estimating functions, but I am not sure whether that would work.

@CarterButts , @handcock , @drh20drh20 , @schmid86 , any thoughts?